Performance of Deep Learning Algorithms for Segmentation of Shoulder Anatomy on MRI: A Systematic Review.
Authors
Affiliations (3)
Affiliations (3)
- Department of Orthopedics and Traumatology, Oncology Unit, Hospital Italiano de Buenos Aires, Buenos Aires, Argentina.
- Department of Orthopaedic Surgery, Teikyo University School of Medicine, 2-11-1 Kaga, Itabashi-ku, Tokyo, 173-8605, Japan.
- Department of Orthopaedic Surgery, Teikyo University School of Medicine, 2-11-1 Kaga, Itabashi-ku, Tokyo, 173-8605, Japan. [email protected].
Abstract
To synthesize the performance and methodological characteristics of fully automated deep learning algorithms for shoulder magnetic resonance imaging (MRI) segmentation. This systematic review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement. CENTRAL, PubMed/MEDLINE, Scopus, Embase, Web of Science, and IEEE Xplore were searched from inception through July 17, 2026. The protocol specified fully automated deep learning-based pixel-wise segmentation of shoulder MRI, an expert-derived MRI-space reference standard, and quantitative segmentation metrics. Two reviewers independently screened records; unresolved disagreements were adjudicated by a senior reviewer. Overlapping datasets were linked during synthesis. Of 354 identified records, 169 duplicates were removed, 185 records were screened, and 53 full-text reports were assessed. Thirty-five reports were included (31 journal articles and 4 full conference papers; 2018-2026). U-Net-based encoder-decoder architectures predominated, with additional region-based, transformer, adversarial, and autoencoder approaches. Dice similarity coefficients ranged from approximately 0.81-0.99 for osseous targets, 0.71-0.99 for shoulder muscles, and 0.58-0.94 for tendon or tear regions; intersection-over-union values for tear or surgical-planning regions ranged from 0.45-0.83. MRI protocols and validation designs were heterogeneous, and most reports described retrospective single-center investigations. Five reports used independent external test sets; several others included multicenter images. Overlapping cohorts precluded a valid pooled participant count. Deep learning achieved high accuracy for shoulder bones and rotator-cuff muscles in selected datasets, whereas tear-region performance was more variable. Multicenter patient-level validation and standardized annotation and reporting are needed before routine clinical deployment.